Fail Closed AI Agent Guardrails: Why Empty Is Not Safe
An empty HOME migrated my real database. An empty git status force-deleted four worktrees. Five fail-closed rules for AI agent checks.
$ grep -r "#ai-agents" ~/blogs # my words, with labels
I tagged my opinions.
Even my tangents have an index. I made it easy to keep up.
$ ls ~/write/connections
my posts, connectedI have receipts for that.
5 posts
These are my #ai-agents posts. I made the connection; you get the links.
How to Run Multiple Claude Code Agents at Once (2026)
Run multiple Claude Code agents at once, plus Codex and Agy, from one terminal: a free one-command herdr setup with 3-keystroke launchers and a fleet CLI.
I Asked an AI Agent to Go Live. It Exposed My Server.
A live test of DeepSeek Harness, a self-hosted AI coding agent, shows what happens when an obedient agent is asked to expose a server, and why the firewall around it mattered more than the model inside it.
How Claude and Codex Learned to Tag Each Other
Block's Buzz gives AI coding agents their own keypairs and channel membership. A code-level walkthrough of how one agent wakes another up, with no orchestrator.
MCP: The USB-C of AI, Explained Properly
MCP is called the USB-C of AI. What that analogy actually means, why AI tool calling needed a standard, and what MCP still does not solve for you.